Chapter 3 · Cognition, Intelligence, and Consciousness: What Sits on Each Side of the Screen

Extended content

Liquid part: this text keeps growing after the printed edition.

Three sentences and a sheet of paper

Before the week is out, write down three claims about AI that you hear said out loud: in a meeting, in a sales email, over coffee. Copy them word for word, because the trick is usually in the verb. Under each one, write its plane: cognition, intelligence, or consciousness. And one question: what proof would this sentence bring, if it had to bring any?

They almost never fit cleanly. "It understands what I ask" sounds like a matter of cognition and drags something of the third plane along with it. Note that friction. It is exactly what the map exists to detect.

A worked example

Silvia coordinates operations at a logistics company, and she has been invited to the meeting where her department's AI usage policy is being drafted. In twenty minutes she hears three things. A director says the assistant "already reviews contracts better than the team." A technician answers that it "just completes text." And Dani, a colleague, admits he feels bad correcting it: you can tell it understands, he says, and it even seems to take it badly.

Silvia doesn't argue. She sorts. The director's sentence mixes cognition and intelligence and lacks the rest of the sentence: better on which clauses, against what unusual contract, at what failure rate, verified by whom. The technician's describes the mechanism and skips what the mechanism produces at scale; it works as a corrective to the hype, but not as a portrait. Dani's belongs to the third plane. Here Silvia takes no side: she says that today nobody can assert it or rule it out, and that the policy doesn't need to settle it. It needs to know who answers for what comes out.

She ends by proposing two concrete things. A ten-minute test using the team's own contracts, and a line in the policy naming who signs off on each output.

A ten-minute test

Claims on the cognitive plane can be checked, and almost nobody checks them. Run the test yourself.

1. Take a document of yours, one you know thoroughly, and plant three errors that only an attentive reader would catch: a figure that doesn't match another, an impossible date, a condition that contradicts itself two pages later.
2. Ask the model to review it, with no hints.
3. Write down what it caught, what it let through, and in what tone it let it through.
4. Repeat with a stranger document, a type you rarely see in your field.

Step four probes the intelligence plane: how it performs against the new. The most instructive part, though, is step three. Notice how the machine sounds when it fails. It sounds the same as when it gets things right.

Questions to work with your map

• Which plane am I sloppiest with when I talk about AI?
• What decision this week would change if the model could feel? What if it were impossible for it to feel?
• Which task have I treated like a search engine, and which like an oracle?
• Who answers for the output I've used most this month, and do they know it?

If your answer to the second one is "none," you have understood why the chapter leaves that door marked and closed.

Two columns, with a date

Take a sheet and draw two columns. On the left, capabilities in motion: what you believe the model can't do today. Date it. On the right, stable absences: accountability, position, lived consequences, and what role each plays in your own work. Keep the sheet.

In three months, go back to the left column and cross out whatever no longer holds. The right column shouldn't have changed. If you've crossed something out there, you most likely mistook a capability for an absence, and that's the costliest mistake of all: expecting from the next model what no model brings.

Moving on

You now know which plane to look at. Chapter 4 takes the map to the workplace: what this alien cognition can bear, what it can't, and what falls to you in the meantime.